Subzero: Use cvttss2si and similar instead of cvtss2si for fp->int casts.

src/IceTargetLoweringX8632.cpp

Index: src/IceTargetLoweringX8632.cpp
diff --git a/src/IceTargetLoweringX8632.cpp b/src/IceTargetLoweringX8632.cpp
index 7b977585b5c6e910cb6bc8f2e02f4b678c518d15..a6c09b12626287390c2db609e91831e2d8665d83 100644
--- a/src/IceTargetLoweringX8632.cpp
+++ b/src/IceTargetLoweringX8632.cpp
@@ -2057,6 +2057,7 @@ void TargetX8632::lowerCast(const InstCast *Inst) {
              Inst->getSrc(0)->getType() == IceType_v4f32);
       Operand *Src0RM = legalize(Inst->getSrc(0), Legal_Reg | Legal_Mem);
       Variable *T = makeReg(Dest->getType());
+      // TODO(stichnot): _cvt or _cvtt?

[jvoung (off chromium), 2014/09/08 17:27:05]

seems like it should be cvtt to match the scalar behavior?

[Jim Stichnoth, 2014/09/08 17:47:06]

You're probably right.  I wasn't sure since there were no vector cross test
failures after adding fractional values to the tests.  But on closer inspection,
it looks like there are no vector cross tests for cast operations. :(

       _cvt(T, Src0RM);
       _movp(Dest, T);
     } else if (Dest->getType() == IceType_i64) {
@@ -2079,7 +2080,7 @@ void TargetX8632::lowerCast(const InstCast *Inst) {
       // t1.i32 = cvt Src0RM; t2.dest_type = t1; Dest = t2.dest_type
       Variable *T_1 = makeReg(IceType_i32);
       Variable *T_2 = makeReg(Dest->getType());
-      _cvt(T_1, Src0RM);
+      _cvtt(T_1, Src0RM);
       _mov(T_2, T_1); // T_1 and T_2 may have different integer types
       _mov(Dest, T_2);
       T_2->setPreferredRegister(T_1, true);
@@ -2114,7 +2115,7 @@ void TargetX8632::lowerCast(const InstCast *Inst) {
       // t1.i32 = cvt Src0RM; t2.dest_type = t1; Dest = t2.dest_type
       Variable *T_1 = makeReg(IceType_i32);
       Variable *T_2 = makeReg(Dest->getType());
-      _cvt(T_1, Src0RM);
+      _cvtt(T_1, Src0RM);
       _mov(T_2, T_1); // T_1 and T_2 may have different integer types
       _mov(Dest, T_2);
       T_2->setPreferredRegister(T_1, true);